AI hardware, explained.
Short, sourced explainers on the parts inside AI systems. Each one is the reading version of a 3D teaching model, so you can switch between the text and the hardware.
What is high-bandwidth memory (HBM)?
How HBM stacks DRAM beside a GPU’s compute dies, why bandwidth matters as much as capacity, how the stacks are made, and where HBM3E and HBM4 fit.
Read the explainerWhat is a GPU compute die?
What a GPU compute die does, why Blackwell Ultra uses two dies and AMD MI300X uses eight chiplets, and what the lithography reticle limit has to do with it.
Read the explainerWhat is a silicon interposer?
How an interposer wires GPU compute dies to HBM inside one package, and how TSMC’s CoWoS-S, CoWoS-R and CoWoS-L approaches differ.
Read the explainerWhat does a GPU package substrate do?
The package substrate supports a GPU’s chips, carries power and signals toward the board, and works with the interposer. Here is how the layers divide the job.
Read the explainerWhat is inside an AI compute tray?
The roles of GPUs, Grace CPUs, network adapters, power delivery and liquid cooling in an NVIDIA GB300 reference compute tray, and how they depend on each other.
Read the explainer
Prefer to start with the models? Explore a data center, a GPU or a compute tray.